| Intended Task/Domain: |
Mesoscale satellite imagery forecasting / Short-range weather nowcasting |
| Model Type: |
Diffusion Transformer (DiT) with 2D neighborhood attention |
| Intended Users: |
Weather researchers, meteorologists, and operational forecasters working with Meteosat FCI data. |
| Output: |
Tensor of 16 FCI multispectral channel radiances representing the predicted satellite state at the next 10-minute timestep. |
| Describe how the model works: |
Six consecutive 10-minute FCI multispectral satellite frames (16 channels each), together with three solar geometry variables and eight static conditioning variables (terrain elevation, water body fraction, and satellite/Earth geometry), are fed through a diffusion transformer with 2D neighborhood attention. The model generates the next satellite state by iterative denoising. Multiple ensemble members can be produced by sampling different noise realizations. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: |
Not Applicable |
| Technical Limitations & Mitigation: |
Predictions may degrade for weather conditions underrepresented in the November 2024–May 2026 training period (e.g., rare or extreme events). Autoregressive error accumulates over longer lead times. |
| Verified to have met prescribed NVIDIA quality standards: |
Yes |
| Performance Metrics: |
Deterministic and Probabilistic Error (RMSE, CRPS); Probabilistic Calibration (rank histograms, spread-to-skill ratio); extreme events metrics (FSS, CSI) |
| Potential Known Risks: |
The model may incorrectly predict cloud structures, convective initiation, or storm intensity. Outputs should not be used as the sole basis for safety-critical operational decisions without independent validation. |
| Licensing: |
Linux Foundation OpenMDW License Agreement, version 1.1 |
